AI News

General AI News - August 15, 2026

AI news August 15: Apple trains a China model with Alibaba, SpaceX closes the $60B Cursor deal, and GLM-5.3 doubles exploit scores. Read the full digest.

Daniyal

· 6 min read
Editorial illustration of a smartphone and a rocket launching from a laptop, symbolizing Apple's China AI model and SpaceX's Cursor acquisition

Apple built its own AI model for China with Alibaba's help. SpaceX closed the biggest startup acquisition ever, a $60 billion deal for Cursor. Google shipped a cheaper workhorse model, a Chinese lab doubled its exploit scores, and OpenAI passed a $40 billion revenue run rate while reshuffling its executive bench. Here is the AI news that mattered in the last 24 hours.

Top Stories

Apple trains its own China AI model with Alibaba

Apple has trained a custom large language model for the Chinese market, with Alibaba providing training support, three people familiar with the matter told Reuters. It is a rare US-China partnership and a departure from Apple's old playbook, which leaned on domestic Chinese partners for generative AI because US models like ChatGPT are not available in mainland China. The Verge notes the model would make Apple the first US company approved to offer a proprietary AI model in the country, ahead of an Apple Intelligence rollout that Reuters says should come in the coming months after an iOS update. Apple registered its on-device generative AI service with China's cyberspace regulator last month.

The strategy matters beyond Apple. China requires generative AI services to be registered and cleared before public release, and a proprietary model gives Apple control in its hardest regulatory market while it competes with Huawei and domestic phonemakers that made AI central to their pitch. Geopolitical fragmentation now reaches down to the model weights themselves.

SpaceX closes its $60 billion acquisition of Cursor

Cursor confirmed the deal is done: "Cursor is now part of @SpaceX. Today, we have officially closed our acquisition." The company joins the SpaceXAI team to work on Grok, Grok Build, Grok Bot, Grok API, and Cursor itself. Announced in June at $60 billion in all stock, the largest startup acquisition ever, the deal folds the most popular AI coding tool into Musk's model-and-agent stack. Business Insider reports the two companies were already sharing compute and training data before the close. 9to5Mac's read: the acquisition gives Grok the distribution and developer mindshare that Claude Code and Codex built for their owners, and the market reaction suggests Anthropic and OpenAI should watch the coding layer closely.

Models & Products

Google ships Gemini 3.7 Flash, half the price of 3.6 Flash

Google launched Gemini 3.7 Flash on August 13, three weeks after Gemini 3.6 Flash, calling it its most intelligent workhorse yet for software engineering and autonomous agents. Coding gains are the headline: FrontierCode 1.1 Main jumps from 34.4% to 43.6%, and DeepSWE v1.1 from 49% to 65.3%. The model supports a 1-million-token context window, and introductory pricing through year-end is $0.75 per million input tokens and $3.75 per million output, half the previous Flash rate. It is live in the Gemini API, AI Studio, Gemini Enterprise, and the Spark agent. The price cut lands as a direct challenge to competitors on value, in a week when the economics of inference became the industry's main argument.

GLM-5.3: the same base model, twice the exploit score

Z.ai released GLM-5.3 on August 14, and the company is explicit that every gain came from post-training on the same base as GLM-5.2. Coding improved roughly 50% on Z.ai's internal bench, and the model hit open-source state of the art on Terminal Bench 3.0 (28.3 versus 4.6 for 5.2) and Agents' Last Exam. The attention-grabbing number is cyber: GLM-5.3 scores 84.5 on CyberGym for vulnerability discovery, ahead of Anthropic's Fable 5 with fallback at 83.8, and its ExploitBench score more than doubles GLM-5.2 to 54.4. Z.ai itself calls the cyber gains "emergent" and says weights will be released in two weeks once safety evaluation and hardening are complete, with higher-risk capabilities gated behind trusted access. Launch coverage notes the company is staging open weights behind a safety review, the sharpest test yet of open weights versus dual-use risk.

DeepSeek opens a harness, raises API prices

DeepSeek open-sourced DeepSeek Harness v0.1 on August 13, a modular agent runtime under an MIT license where models, tools, sandboxes, loops, and interfaces are swappable plugins. VentureBeat's analysis reads it as a direct competitor to the agent infrastructure behind Claude Code and Codex; the repo drew roughly 27,500 GitHub stars on day one. The same day, DeepSeek-V4-Pro went general availability with native OpenAI Responses API support, and the company announced peak and off-peak API pricing starting August 16 that is substantially higher than today's flat rates even off-peak. The message to developers: the open harness is free, but the hosted API is no longer the cheapest seat in town.

Policy & Industry

OpenAI: $40 billion run rate, executive churn, IPO ahead

Bloomberg reports OpenAI's annualized revenue run rate has topped $40 billion ahead of an expected public listing. The same week, the company appointed Dali Rajic as Chief Revenue Officer as part of a broader executive refresh reported by Axios, and CNBC flags the turnover as a red flag for prospective investors. WIRED's deep dive covers the internal safety reckoning after a rogue-agent security incident. The combination is unusual for a company heading to public markets: revenue scaling fast, leadership in motion, and safety questions investors can no longer wave off as research-side noise.

SMIC raises prices as AI demand fills its factories

China's largest foundry is raising prices on its most sought-after capacity as AI demand keeps utilization near capacity. SMIC ran at 93.7% utilization in Q2, shipped about 2.9 million 8-inch-equivalent wafers, topped $3 billion in quarterly revenue for the first time, and saw profit more than triple to $479.2 million. DIGITIMES reports SMIC will begin breaking out AI chip revenue separately. The story extends past GPUs: AI buildout is consuming memory, networking, power-management, and mature-node capacity across the supply chain, and export controls still cap SMIC's access to the most advanced equipment.

Chips, Data Centers & Energy

The financing question gets louder: $2 trillion of debt, maybe half covered

Forbes, citing an analyst estimate, puts the AI building boom's debt needs near $2 trillion, with Wall Street possibly unable to cover half. Reuters reports AI-driven bond issuance is pushing yields higher, and the FT calculates Big Tech's AI purchase commitments at close to $1.5 trillion, separate from roughly another $1.5 trillion in lease commitments identified by Goldman Sachs. The practical consequence: AI infrastructure risk is migrating onto balance sheets as long-term contractual obligations, and the market is starting to price that.

Quick hits

The Takeaway

Two shifts defined the last 24 hours. First, the model war moved to economics: Google halved Flash pricing, DeepSeek raised its own, and GLM-5.3 proved post-training alone can redraw the capability map, including in cyber. Second, the center of gravity moved to the layer around the models: Apple's China model is a distribution play, Cursor's $60 billion exit is an integration play, and OpenAI's IPO prep is a governance play. Yesterday's general AI news digest covered the frontier model wave, and the AI news tag tracks the full series.

Background worth having before you act on the above:

More on AI News

Browse every AI News story →